Abstract Phishing is the technique of stealing personal and sensitive information from an email client by sending emails that impersonates like the ones from some trustworthy organizations. Phishing mails are a specific type of spam mails; however the effects of them are much more terrible than alternate sorts. Mostly the phishing attackers aim the clients of the financial organizations, so its detection needs high priority. Lots of research activities are done to detect the phished emails, in the proposed methodology a multi-classifier prediction model is introduced for detecting phished emails. Our contention is that solitary classifier prediction might not be satisfactory to urge the clearest picture of the phishing email; multi-classifier prediction has accuracy 99.8% with an FP rate of 0.8%. Keywords: Phishing email detection, Machine learning techniques, Multi-classifier prediction model, Majority voting

1. IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 03 Issue: 03 | Mar-2014, Available @ http://www.ijret.org 31
A MULTI-CLASSIFIER PREDICTION MODEL FOR PHISHING
DETECTION
Sarju S1
, Riju Thomas2
, Emilin Shyni C3
1, 2
PG Scholar, 3
Associate Professor, Department of Computer Science and Engineering, KCG College of Technology,
Tamilnadu, India
Abstract
Phishing is the technique of stealing personal and sensitive information from an email client by sending emails that impersonates like
the ones from some trustworthy organizations. Phishing mails are a specific type of spam mails; however the effects of them are much
more terrible than alternate sorts. Mostly the phishing attackers aim the clients of the financial organizations, so its detection needs
high priority. Lots of research activities are done to detect the phished emails, in the proposed methodology a multi-classifier
prediction model is introduced for detecting phished emails. Our contention is that solitary classifier prediction might not be
satisfactory to urge the clearest picture of the phishing email; multi-classifier prediction has accuracy 99.8% with an FP rate of 0.8%.
Keywords: Phishing email detection, Machine learning techniques, Multi-classifier prediction model, Majority voting
----------------------------------------------------------------------***------------------------------------------------------------------------
1. INTRODUCTION
Nowadays the emails turn into one of the generally utilized
communication medium within the globe. Because of the fame
of emails, the attackers utilized it to snatch the client data.
Phishing messages are a particular kind of spam mail, which is
used to take the individual and fiscal data from the email
clients. Generally the attackers send an email that looks like
legitimate messages from some reputed organizations, which
lead clients to phishing sites. Phishing sites always have a user
entry form, when he enters his data like user names,
passwords and credit card details which in turn utilized by the
attackers to do some deceitful exercises. As per the latest
report issued by APWG [1] (Anti-Phishing Working Group),
amount of phishing attack evolved and burgeoned in
overabundance of 20% in 2013. Latest Trends Report of
APWG quotes that the overall number of distinctive phishing
internet sites rose to 143,353 throughout the July-September
that is over the past quarter's 119,101.
Heaps of researches are carried out to detect the phished mails,
in which the machine learning techniques are most prominent
one due the higher precision of detection. The classification
algorithms developed in the learning techniques are utilized
for anticipating the class of the given mail, but each
classification algorithms have their own particular blemishes.
In the proposed technique we implemented a multi-
classification method which fuses three most accurate
classifiers for foreseeing the class label. A prediction model is
constructed with the classifiers which incorporate Support
Vector Machine (SVM), J48 and Instance Base 1, majority
voting algorithm is used to settle on the last choice in regards
to the class name of the given email. The dataset holds 5260
publically accessible email corpus for train the prediction
model and 500 messages are utilized as the test mails.
2. BACKGROUND
Recently, a variety of anti-phishing techniques are introduced,
out of which machine learning technique based approaches are
most prominent one. Features are extracted from the html part
and body part of the email is used by the machine learning
algorithms to predict the possibility of the phishing.
Chandrasekaran et.al [2] proposed phishing email detection
based on the structural properties of the phished mails. A total
of 25 structural features are extracted and classification of the
mails is done using the Support vector Machine algorithm.
Data set only contains 200 mails from the public phishing mail
corpus, so the results might not be accurate. Sarju et al.[3]
utilized the structural properties to detect the spam emails and
used Naïve Bayes, Adaboost and Random Forest are used to
measure the accuracy. From the above works it identified that
the structural properties can be used to discriminate the
phished emails from ham mails.
A number of other novel features are also useful to identify
phished emails. Bergholz et.al [4] analyzed different
properties of the email to detect the phished ones. The content
of the emails are evaluated for constructing the feature set and
is used for phishing detection. Random Forest and SVM are
used to measure the accuracy of the detection.
Abu-Nimeh et al.[5] compared machine learning techniques in
phishing detection, they used six different classifiers. A total
of 43 features are extracted from the dataset of 2889 phishing
and ham emails. They found that Random Forest outperforms

2. IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 03 Issue: 03 | Mar-2014, Available @ http://www.ijret.org 32
all other classifiers used in their methodology. Miyamoto et al.
[6] analyzed nine machine learning techniques for detection of
phishing sites; they found that Random Forest and SVM
outperforms all the other classifiers.
3. PROPOSED METHOD
In the proposed method structural features, content based
features and element features of emails are analyzed and
extracted for training the prediction model as shown in the
Figure 1, Fi represents the feature set extracted from the mails
and C1 and C2 the corresponding class labels.
Fig 1 - Multi-Classifier Prediction system for Phishing email
detection
3.1 Feature Extraction & Analysis
Feature Extraction and analysis phase, extracts the different
features that plays important role in detecting the phished
emails. In this work, we extracted the structural features,
content based features and element features from the emails.
Emails are available in the Multipart Internet Mail Extension
(MIME) format, an HTML parser is used to parse the mail and
a HTML is tree constructed. Structural features are extracted
from the HTML tree and are multi part count, non ASCII
characters, non-text content and content labeling. The
multipart feature divides the email into different parts. For
example an email with an attachment has two parts, text
content and attachment. An email in the MIME format
contains a header which shows the character encoding used in
that mail. This can be used to identify whether any non ASCII
characters used in that mail. Table 1 shows the content label
types used in this paper.
Table 1 - Email Content Types
Element features includes the web technologies used in the
email. In this work, we extracted element features of type
Boolean which indicates whether HTML, JavaScript, VB
Script, XHTML, and CSS used in the mail. Finally the content
based features available in the email are extracted. Totally 42
features are extracted from the email corpus and give it as a
training set to the prediction model.
3.2 Prediction Model
The extracted features from the feature extraction and analysis
stage are used in the multi-classifier prediction model. The
prediction model is built using the machine learning
algorithms which includes J48, SVM and IB1, each one is
capable of classifying the mails into phished or ham mails.
The accuracy of the prediction can be improved by combining
the classifiers. The final decision regarding the category of the
mail is done through the majority voting algorithm. Different
research works are done to combining classifiers [7-9].
Majority Voting is the mostly used way for combining
classifiers, which count the votes for each class that are
predicted by the classifiers and majority class is selected. The
new confidence fi x for class i is calculated as
fi x = I(maxj pji x = j)j (1)
in which I() is the identifier function: I(x) =1 if x is true else
I(x) will be zero.
When a new mail comes the prediction model identifies its
category based in the training test given.
4. EXPERIMENTAL EVALUATION
The dataset holds 5260 publically accessible email corpus of
phished and ham mails for train the prediction model and 500
messages are utilized as the test mails. The performance of the
prediction model is analyzed using different measures like
True Positive, False Positive, Accuracy and Receiver Operator
Characteristics curve.
File Extension MIMEType Description
.txt text/plain Plain text
.htm text/html Styled text in HTML format
.jpg image/jpeg Picture in JPEGformat
.gif image/gif Picture in GIF format
.wav audio/x-wave Sound in WAVE format
.mp3 audio/mpeg Music in MP3 format
.mpg video/mpeg Video in MPEGformat
.zip application/zip Compressed file in PK-ZIP format

3. IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 03 Issue: 03 | Mar-2014, Available @ http://www.ijret.org 33
The information about actual and predicted classifications
done by machine learning systems is represented in the form
of a confusion matrix [10] as shown in the figure 2 and
accuracy is measured based on entries.
Fig 2 - Confusion Matrix
Accuracy =
TP + TN
TN + FN + FP + TP
(2)
If the mail is ham and it is classified as ham, it is counted as a
True Positive (TP); if it is classified as phish, it is counted as a
False Negative (FN). If the mail is phished and it is classified
as phish, it is counted as a True Negative (TN); if it is
classified as ham, it is counted as a False Positive (FP).
The Figure 3 compares the FP rate obtained when the
classifiers used independently and also combined using
majority voting. It is shown that the multi-classification using
majority voting outperforms individual classifier performance
with an FP rate of 0.8%
.
Fig 3 - Phishing Prediction FP Rate
The accuracy of the prediction model is evaluated using the
equation 2. From the results, it is clearly understood that the
accuracy of the multi-classification is higher than the
classifiers applied individually. Multi-classification with
majority voting gains an accuracy of 99.80% and is shown in
the Figure 4.
Fig 4 - Phishing Prediction Accuracy Comparison
Another performance measure used to evaluate our proposed
system is ROC [11] curve. In an ROC graph X axis plotted
with FP rate and TP on Y axis. Figure 5a and 5b shows the
ROC measures for the prediction model, by analyzing the
graphs it is identified that the multi-classification prediction
model has improved results compared to the independent
classifier prediction model.
(a)
(b)
Fig 5 – ROC Curves (a) when classifiers used independently ;
(b) Multi- classification using Majority Voting
Phish Ham
Phish TN FN
Ham FP TP
Predicted
Actual

4. IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 03 Issue: 03 | Mar-2014, Available @ http://www.ijret.org 34
5. CONCLUSIONS
Our argument is that single classifier would not be adequate to
urge the clearest image of the phishing email detection
accuracy. The experimental results shown that proposed using
multi-classifier prediction model outperforms the individual
classifier based prediction models in many aspects. It
preserves an accuracy of 99.8% with an FP rate of 0.8%. The
ROC measure shows that the multi-classification with
majority voting gives almost an ideal curve compared to
independent classification algorithms.
In the future work, we are planning to incorporate the topic
modeling features to the feature set generation stage, because
it has the capability to overcome the novel techniques used by
the attackers.
REFERENCES
[1] APWG, Anti phishing working group.
http://wwww.antiphishing.org (2013). Accessed 31
September 2013
[2] Chandrasekaran M, Narayanan M and Upadhyaya S.:
Phishing email detection based on structural
properties. In: Proceedings of 9th Annual NYS Cyber
Security Conference, pp. 2-8 (2006).
[3] Sarju S, Riju Thomas and Emilin Shyni C.: Spam
Email Detection using Structural Features.
International Journal of Computer Applications
89(3):pp.38-41 (2014).
[4] A Bergholz, JH Chang, F Reichartz, and S Strobel.:
Improved Phishing Detection using Model-Based
Features. In Proceedings of the Conference on Email
and Anti-Spam (CEAS), Mountain View, CA,(2008).
[5] Saeed Abu-Nimeh , Dario Nappa , Xinlei Wang , Suku
Nair.: A comparison of machine learning techniques
for phishing detection, In. Proceedings of the anti-
phishing working groups 2nd annual eCrime
researchers summit, pp.60-69, (2007).
[6] Daisuke Miyamoto , Hiroaki Hazeyama and Youki
Kadobayashi.: An evaluation of machine learning-
based methods for detection of phishing sites,
Proceedings of the 15th international conference on
Advances in neuro-information processing, (2008).
[7] Geman D and Geman S.: Stochastic relaxation, Gibbs
distribution, and the Bayesian restoration of images.
IEEE Trans. on Pattern Analysis and Machine
Intelligence, PAMI-6(6), pp.721–741, (1984).
[8] Jain AK, Zhong Y and Lakshmanan S.: Object
Matching Using Deformable Templates. IEEE Trans.
Pattern Analysis and Machine Intelligence 18(3),
pp.267–278, (1996)
[9] Kumazawa I.: Shape extraction by cellular Hough
transform, Technical report of IEICE, PRMU, pp.96–
105, (1996)
[10] Provost F, Fawcett T and Kohavi R.: The case against
accuracy estimation for comparing induction
algorithms. In. Proceedings. ICML-98.pp. 445–453,
(1998)
[11] T. Fawcett, An Introduction to ROC Analysis, Pattern
Recognition Letters 27(8), pp. 861-874, (2006).
[12] WEKA. http://www.cs.waikato.ac.nz/ml/weka/ (2013):
Accessed 31 November 2013
[13] SpamAssassin, http://spamassassin.apache.org (2013)
: Accessed 31 November 2013
[14] Phishingcorpus http://monkey.org (2013), Accessed 31
November 2013.
[15] Apache James. (2013) Mime4J Parser.
http://james.apache.org/mime4j: Accessed 14
October 2013.